Local, Quantitative Morphometry of Fibroproliferative Lung Injury Using Laminin

Abstract

Investigations into the mechanisms of injury and repair in fibroproliferative disease require consideration of the spatial heterogeneity inherent in the disease. Most scoring of fibrotic remodeling in preclinical animal models relies on the modified Ashcroft score, which is an ordinal rubric of macroscopic resolution. The obvious limitations of manual histopathologic scoring have generated an unmet need for unbiased, repeatable scoring of fibroproliferative burden in tissue. Using computer vision approaches on immunofluorescence imaging of the extracellular matrix component laminin, we generated a robust and repeatable quantitative remodeling scorer. In the bleomycin lung injury model, the quantitative remodeling scorer shows significant agreement with the modified Ashcroft scale. This antibody-based approach is easily integrated into larger multiplex immunofluorescence experiments, which we demonstrate by testing the spatial apposition of tertiary lymphoid structures to fibroproliferative tissue, a poorly characterized phenomenon observed in both human interstitial lung diseases and preclinical models of lung fibrosis. The tool reported in this article is available as a stand-alone application that is usable without programming knowledge.

Interstitial lung diseases constitute a diverse group of lung pathologies with various drivers of tissue injury that ultimately lead to aberrant tissue remodeling and scarring. Idiopathic pulmonary fibrosis is an interstitial lung disease typified by chronic, progressive interstitial fibrosis with a characteristic histologic appearance (1). Loss of lung function caused by fibroproliferation results in a mean survival time of 3 years after diagnosis. Histopathologic analyses of idiopathic pulmonary fibrosis, presenting as usual interstitial pneumonia, are a core component of ongoing research into the mechanisms driving fibrotic disease. The pathologist’s diagnosis of usual interstitial pneumonia as described by guidelines updated in 2022 requires 1) normal lung parenchyma punctuated with regions of fibrotic remodeling; 2) fibroblastic foci (myxoid, pale staining subepithelial fibroblasts); 3) the accumulation of extracellular matrix (ECM), or scar; and 4) honeycombing (dilated cystic spaces lined by airway epithelium located within mature scar), in addition to the absence of features that suggest other diagnoses (2–4). These fibroblastic foci can be observed at the boundary between uninvolved tissue and advancing scar and are where tissue remodeling and scar-generating myofibroblasts proliferate (5). Accounting for this spatially variable biology is, therefore, an important factor in the study of pathologic mechanisms (6, 7). Among these focal features, aggregates of lymphocytes in the lung parenchyma have been observed clinically and in sterile preclinical models of lung fibrosis (8–11). The degree to which these tertiary lymphoid structures (TLSs), frequently also described as induced bronchus-associated lymphoid tissue, are participatory in the pathology of fibrosis is largely unknown. Adequately phenotyping these phenomena in preclinical animal models is an evergreen problem for fibrosis researchers.

The Ashcroft scoring system, put forward by Ashcroft and colleagues in 1988 (12) and modified further by Hübner and colleagues in 2008 (13), is an expansion on even earlier disease-staging guidelines for asbestosis (14), and consists of eight grades of fibrosis and a score of 0 for healthy lung. The scoring is tacitly performed on lung parenchyma, with contributions from airways and larger blood vessels excluded heuristically by pathologists. Grades 1 through 4 are characterized by the thickening of interstitial tissue but no loss of alveolar architecture. Grades 5 through 8 include distinct fibrotic masses and progress in severity with the proportion of parenchyma displaced by expanding interstitium up to complete ablation of normal tissue at a score of 8. This scoring is traditionally performed on sections stained with Masson’s trichrome, which emphasizes collagen content, but this is not required.

For several reasons, the modified Ashcroft system is not conducive to multiplexed imaging approaches. It requires histochemical staining, which is generally incompatible with multiplex immunofluorescent imaging. It is traditionally scored on large fields of view, limiting the resolution of fibrotic scoring to gross evaluations. Finally, it is performed manually, introducing bias and variability. There are computational methods in active development which automate the histopathologic scoring of chemical stains, successfully removing the burden of labor and interobserver/intraobserver variability from the analysis, but these methods do not provide the capability to multiplex with other targets (15, 16). Immunofluorescence imaging benefits from target specificity and ease of multiplexing but is generally deficient in broader tissue context, as only targeted antigens are visible. Frequently, ECM proteins such as collagens, laminins, or elastins are targeted in multiplex immunofluorescent panels to provide this missing structural information (17).

A distinct structure in healthy tissue is the basement membrane, a thin sheet of ECM on which cells are anchored and derive their polarity. To maximize gas exchange in the lung, the alveolar basement membrane is quite thin and is frequently shared between the capillary endothelium and alveolar epithelium. The primary components of this basement membrane by weight are laminin and collagen IV. The loss of organized basement membrane structure is thought to mark irreversible lung injury; functional tissue has been remodeled by fibroproliferative processes such that the blueprint for healthy tissue has been lost (18). Although laminin and/or collagen IV are frequently targeted for immunostaining, there is an unexploited opportunity to extract further information about tissue health from these basement membrane proteins.

The tool described in this study uses analyses of immunofluorescent signals from a polyclonal antilaminin antibody to compute a local quantitative remodeling score (QRS). The morphological progression of the laminin network during fibroproliferation can be described as a broadening of the cross-sections of alveolar basement membrane and a loss of contrast between basement membrane and adjacent tissue. These qualitative phenomena are found concomitant with reduced alveolar air space and expanding interstitial space in the tissue. These morphological changes are easily visible to the human observer but require textural image analysis to be quantified, a field of image processing common in medical and topographic/satellite imaging. Herein, we validate our tool against the common histopathologic scoring rubric of the modified Ashcroft system in the bleomycin preclinical animal model and test the association of TLSs with fibroproliferative tissue as a novel analysis enabled by this method.

This article was previously published in preprint form (www.biorxiv.org/content/10.1101/2023.06.15.545119v1).

Methods

For a troubleshooting guide and full tutorial for the QRS app, see the data supplement.

Preclinical Animal Model of Bleomycin Lung Injury

Eight- to 12-week-old C56BL/6J mice (The Jackson Laboratory) were anesthetized in accordance with protocols that were approved by the Institutional Animal Care and Use Committee. Clinical-grade bleomycin sulfate (Pfizer/Fresenius Kabi through the University of Virginia Health System Pharmacy Services) was introduced orotracheally and aspirated at a dose of 1 U/kg in a volume of 50 μl sterile saline. To capture fibroproliferation, mice were killed at 2 weeks post–bleomycin treatment.

Sectioning and Histology of Lungs

Mice were perfused, and their lungs were lavaged once with saline and then inflated with 1% low–melting temperature agarose (Invitrogen) in 1× PBS. Lungs were fixed in 4% paraformaldehyde in 1× PBS for 30 minutes and then cryoprotected in 30% sucrose in 1× PBS until sinking. Lungs were dissected into lobes, embedded in optimal cutting temperature compound, and frozen. The left lobe was sagittally sectioned at a 10-μm thickness onto Superfrost Plus slides (Fisher Scientific).

Histochemistry, Immunostaining, and Imaging

Formalin-fixed, paraffin-embedded (FFPE) sections were treated with proteinase K (Dako S3020) for 5 minutes before the subsequent steps. Sections were permeabilized and blocked as described previously (19). Anticol1a1 (AB_2904565; Cell Signaling Technology), antilaminin (AB_10001146; Novus), anti-CD45R/B220 (AB_2896201; eBioscience) for B cells, and DAPI or NucSpot Live 488 (Biotium) were used for nuclear staining. Stained samples were mounted in SlowFade Glass (Invitrogen). Micrographs were acquired on a Leica Thunder total internal reflectin fluorescence, or TIRF, instrument in epifluorescence (Leica DMi8), using a 20× objective (Leica HC PL APO 20× /0.80 DRY) and a DFC9000 GTC scientific complementary metal–oxide–semiconductor, or sCMOS, camera with no binning. Laminin fluorescence was corrected for flatness of field using a dyed slide reference and stitched using LAS-X software (Version 3.7.5.24914; Leica Microsystems). For sequential staining of sections, initial immunostaining and imaging were performed as described, and the samples were decoverslipped and then stained for collagen (Masson’s trichrome/aniline blue).

Image Analysis

For the stand-alone application TIFF images, were batch processed with the parameters listed in the data supplement. Images were loaded into QuPath and overlaid if necessary. For modified Ashcroft scoring, 10 random fields of 0.9mm on a side were taken from each whole-slide Masson’s trichrome scan, with four blinded scorers scoring all fields independently in a single session. For the correlation of computed features to modified Ashcroft scores, the mean computed features within the field were compared. (For a full list of correlated features, see the data supplement.) For all immunofluorescence analysis and visualization, the true signal was determined using unstained and secondary-only controls. The laminin channel was used with the thresholding function in QuPath to generate masks of lung tissue. TLSs were manually identified, and distance maps between TLSs and tiles were generated using the QuPath function distanceMap. Measurements were exported, with subsequent manipulations and statistical tests performed in GraphPad Prism 9.0.

Results

Basement Membrane Remodeling in Bleomycin Model of Fibroproliferative Lung Disease

The orotracheal bleomycin injury model for fibroproliferative disease induces collagen deposition, myofibroblast activation and proliferation, and destruction of functional lung parenchyma (20, 21). At 14 days post–bleomycin administration, the acute inflammatory phase is resolved, and instead, fibroproliferation predominates, with maximal scar tissue appreciable at 21–28 days postinjury. Here, we look at early tissue remodeling with the goal of identifying features associated with fibroproliferation before the generation of mature scars.

Heterogeneity of injury phenotype is apparent within a single section at low magnification (Figure 1A) with fibroproliferative (gold box) and healthy (cyan box) tissue delineated and shown in greater detail in Figure 1B. The deposition of collagen in parenchymal space visualized blue by Masson’s trichrome (Figure 1A, left) or with anticollagen antibody (Figure 1A, center) indicates areas of fibroproliferation. Immunostaining for collagen alone does not distinguish between preexisting healthy collagen, as it is abundant in the cuff spaces of healthy airways (Figure 1B, asterisks) and pathologically deposited in the interstitium in fibrosis. The addition of laminin immunostaining provides structural context (Figures 1A and 1B, right), and alveolar spaces become readily identifiable. Laminin alone is sufficient to qualitatively identify tissue remodeling in lung parenchyma: Healthy networks of alveolar basement membrane exhibit a characteristic thin and highly networked structure, whereas the same basement membrane network becomes broad and loses contrast in regions of remodeling (Figure 1C, top and bottom, respectively).

Figure 1.

Basement Membrane Remodeling Phenotype in Bleomycin Lung Injury. (A) Representative micrographs of serial sagittal sections of a left mouse lung 2 weeks post–bleomycin treatment (left). Histochemical staining for Masson’s trichrome and immunofluorescence (middle) for collagen I (anti-Col1a1) alone and (right) with laminin (anti-laminin). (B) Expanded insets of micrographs, taken from (top) healthy and (bottom) fibroproliferative regions of the same section. Asterisks indicate the collagen-rich, laminin-negative cuff space surrounding large airways. (C) Higher resolution fluorescence micrographs of laminin from a serial section of the same lung. Field from (top) healthy and (bottom) fibroproliferative regions of lung parenchyma. Scale bars, 100 μm.

Generation of Computed Laminin Features and Association of Features with Modified Ashcroft

Sequential staining of a single section allows for the 1:1 alignment of multiple imaging modalities performed in series. Initial processing for immunofluorescence imaging is performed as normal. After fluorescence imaging, the sample is decoverslipped, processed for Masson’s trichrome staining, and imaged in color through brightfield microscopy (the workflow is schematized in Figure 2A). Each lung section has been imaged twice: by laminin immunofluorescence and color Masson’s trichrome.

Figure 2.

Validation of the Quantitative Remodeling Scorer (QRS): Association of Computed Laminin Features to Histopathologic Ground Truth. (A) Diagrammatic representation of the lab workflow for sequential staining of a single section for laminin and, subsequently, Masson’s trichrome. (B) Two scoring modalities performed on the immunofluorescence and brightfield acquisitions, respectively. Top: Immunofluorescence scoring is performed on subsampled tiles of the larger image and consists of multiple independent computational measures on per-pixel values and multipixel relationships within each tile. Tiles are color-coded yellow on the basis of their score, with higher yellow values indicating higher scores. Bottom: Histopathologic scoring by using modified Ashcroft is performed on simulated 10× FOV. Each field is given a mean score from four independent, blinded scorers. (C) Top: An overlay of the brightfield trichrome image with the computed laminin feature map. Bottom left: Volcano plot of the r and P values for the Spearman correlation testing on the various computed laminin features against the mean Ashcroft score. Bottom right: The simple linear regression of the highest scoring laminin feature chosen for the QRS, MPW. n = 10 lungs, with 9 or 10 random fields selected per lung (98 fields total) for histopathologic scoring and subsequent comparison and regression analysis. FOV = fields of view; MPW = mean peak width.

The scoring of images across modalities is demonstrated in Figure 2B. Image processing of laminin immunofluorescence proceeds as follows: The image is subsampled into tiles, and laminin feature scores are computed on each tile independently. Feature scores are defined as a singly parametrized analysis, such as gross fluorescence intensity, computed image textures including several Haralick features, and multiple line profile/histogram analyses. A chosen feature is visualized by applying a lookup table to a normalized range of feature scores and generating a heatmap from those tiles. In this way, additional channels are directly appended to the original fluorescence image, each representing a single feature score. Examples of these computed heatmaps can be seen at the top of Figure 2B. For traditional histopathologic scoring, the modified Ashcroft scoring system was used. Simulated single fields of view (∼0.81 mm²) were shown to trained and blinded scorers. The final histopathologic score for each field is the mean of the scorers, as shown at the bottom of Figure 2B.

The two imaging modes and their scores are then compared in Figure 2C. A superimposition of those measures can be seen at the top of Figure 2C. The modified Ashcroft score of all the scored fields were tested against every computed laminin feature. This identified many significant computed features (Figure 2C, bottom left). The mean peak width (MPW) laminin feature was found to have the highest agreement, with a Spearman coefficient (r) of 0.768 (P = 10⁻¹⁹) (Figure 2C, bottom right). (For a full list of the computed features, see Supplemental Table E1.) The MPW serves to measure the thickness of the interstitial basement membrane itself and ignores empty space. A linear regression of the MPW against the modified Ashcroft score across all scored fields of view is significant, with an R² of 0.534 (P < 0.0001). The QRS that is referred to later in the article uses MPW as its scoring metric.

To test the batch-to-batch reproducibility, we treated two cohorts of mice (ns = 6 and 4) with bleomycin, with starting points and endpoints falling on separate days. The agreement of MPW with the modified Ashcroft score across 10 randomly chosen fields per lung, averaged within each independent cohort, were not significantly different and shared a best-fit line (see Figure E1). Finally, a validation cohort of lungs from mice (n = 6) at 4 weeks post–bleomycin treatment was stained and scored as described earlier (see Figure E2). These lungs were embedded in FFPE, with enzymatic antigen retrieval required for laminin immunofluorescence. The correlation between MPW and modified Ashcroft score for this cohort was again significant (P = 0.0003).

Example Use Case: Association of TLSs with Tissue Remodeling

To demonstrate an analysis enabled by QRS’s granular profiling of tissue remodeling, we chose to investigate the association of lymphocytic aggregates, frequently termed TLSs, with fibroproliferative tissue in the bleomycin injury model. B cells were identified in tissue by B220 (CD45R) staining in a multicolor immunofluorescence assay. A representative micrograph can be seen in Figure 3A. TLSs, defined by densely nucleated structures of B cells longer than 100 μm along a long axis, were observed in 9 out of 13 bleomycin-treated mouse lung sections at 14 days post–bleomycin treatment. These nine sections were carried forward into subsequent analyses. A QRS map was computed for whole lung sections, with a resolution of 75 μm. A heatmap of QRS on a lung section can be seen at the top of Figure 3B. A distance field to the nearest TLS was computed and can be seen at the bottom of Figure 3B. To control for the area of tissue (and, therefore, the number of tiles) at increasing given radii from TLSs being unevenly distributed across lungs (see Figure E3), distances were binned into 25-μm intervals, and the average QRS at each radius was plotted (Figure 3C). The continuous QRS was binned as “healthy” or ‘‘fibroproliferative’” with reference to the linear regression to the modified Ashcroft score described in Figure 1 (Figure 3D). With this categorical variable, we then plotted the proportion of tissue scoring as fibroproliferative, calculated as the quotient of fibroproliferative tissue area to total tissue area at a given radius (Figure 3E). A final analysis found significant differences in the median distance of fibroproliferative regions of tissue to TLSs when compared against healthy regions of tissue from the same section (Figure 3F).

Figure 3.

Quantitative Remodeling Scorer (QRS) Enables Testing of Spatial Hypotheses: Fibroproliferative Tissue Associates with Tertiary Lymphoid Structures (TLSs). (A) Representative micrograph of multicolor immunofluorescence panel plus QRS heatmap. Nuclei (NucGreen), B cells (anti-B220), laminin (anti-Laminin), and QRS heatmap (computationally derived from laminin) are visible as (left) single channels and (right) composite overlay. White triangles indicate tertiary lymphoid structures. Scale bars, 100 μm. (B) Parametrized tiles visualized using standard cytometric software, with the field from (A) bounded in red. Top: heatmap of whole-section QRS. Bottom: distance field indicating distance from any tile to the nearest TLS. (C) Plot of the mean QRS at a given distance to the nearest TLS (in 25-μm bins). Individual colored lines represent individual lungs; black shaded line indicates the overall mean ± SEM. (D) Histogram of tissue QRS, with regions gated as “Healthy” and “fibroproliferative” at a QRS of 42. (E) Plot of the fraction of all tissues scoring fibrotic (QRS > 42) at a given distance from the nearest TLS. Individual colored lines represent individual lungs; black shaded line indicates the overall mean ± SEM. (F) Scored lung sections split into “healthy” (QRS < 42) or “fibroproliferative” (QRS > 42). Plotted median distance of “healthy” or “fibroproliferative” tissue from nearest TLS. Scored lung sections are paired. Bars indicate the mean within each group. n = 9, Wilcoxon matched-pairs signed-rank test. *P < 0.05, **P < 0.01, ***P < 0.001. Pos = position.

Discussion

The study of disease is undergoing a revolution of “spatial”-omics. Whereas in the past, highly multiplexed data would require tissue dissociation, the current proliferation of methods enables massively parametrized assays to function in situ, preserving essential spatial context. However, and likely because of methodologic inertia from cytometric techniques, these methods tend to exclude any analysis of ECM morphology or composition. In short, there is a disparity between our capacity to collect data with modern multiplex imaging and our ability to extract useful, testable data from it. This project focuses on one aspect of that disparity: the ability to quantitatively define the level of tissue remodeling across large immunofluorescent acquisitions. We compare this algorithm, QRS, favorably against the current gold standard for quantitatively scoring pulmonary fibrosis in experimental models.

In taking the mean of many line profiles drawn through a tile, an average thickness of interstitial space in that tile can be determined. A benefit of this approach is its inherent robustness to differentials in processing; poorly inflated or locally underinflated regions of lung are differentiated from genuine tissue remodeling so long as the thin laminae of healthy basement membrane are distinguishable. Similarly, peak width does not rely on absolute fluorescence intensity, provided that the laminin signal is sufficiently above the noise floor. Widefield imaging fit for task provided that the tissue thickness is less than the depth of field for the imaging configuration, as shown in this article. As always, consistent processing, staining, and the inclusion of fluorescence controls are required for any analysis. Direct comparisons between tissue processed differently should not be undertaken without consideration of batch correction. Our validation cohort demonstrated a weaker agreement from the initial cohorts (R² 0.13 vs. 0.53). These differences could be a result of differential processing (FFPE vs. cryopreserved), differential section thickness, or, indeed, a different relationship computed between basement membrane texture and modified Ashcroft scoring at the 4-week timepoint as compared to the 2-week. (A full table describing potential issues and solutions in the use of this tool is provided in the supplement. (see Supplemental Table E2.)

Future enhancement of the tool includes a focus on optimizing computational run time and integration into existing histopathologic tools. Certain features that scored highly in our agreement testing, such as the Haralick feature informational measure of correlation 2, are already implemented in software packages such as Qupath (22) and can be computed much more quickly than the mean peak width feature used in the analyses described earlier. It is up to the user to weigh software familiarity and interoperability with existing workflows against the strength of association between laminin features and the modified Ashcroft rubric. Furthermore, the concept of parametrizing healthy tissue ECM structure to identify areas of pathology could be realized through other biomedical image analysis methods. Network analyses could identify where well-defined parenchyma is dysregulated by failing to identify the regular branched architecture in those regions. More complex computational methods, such as computational neural networks, may provide more granular phenotyping of tissue with the potential downside of methodologic opacity.

This tool addresses the growing need for high-throughput computational morphometric methods in lung histology. It provides reproducible, continuous quantification of tissue remodeling at a user-defined spatial scale, reducing the labor and subjectivity involved in human-scored systems. Immunofluorescence multiplexing enables the testing of hypotheses by dissecting the spatial heterogeneity of biology in fibroproliferative lung disease. In this case, we have identified a significant spatial association of tertiary lymphoid structures to early-stage fibrosis in a sterile bleomycin injury model (14 days postinjury). The apposition of these structures to diseased tissue invites further investigation into the potential paracrine and juxtacrine signaling between myofibroblasts and these recruited lymphocytes. More generally, the process of using textural image analysis to identify patterns of healthy and pathologic ECM could, in theory, be extended to other tissues in which repeating, isotropic structures (e.g., alveoli) are perturbed.